Forum poweruser forum

[The year is 2031. Unscoped AI development has been frozen at 2029 levels by an international treaty. The growing awareness of the abstract general problem of the abject state of group rationality spurs the creation of a new kind of forum—by, of, and for forum powerusers—aimed at hot-swapping collective sensemaking structures until a good one is developed.]

0722 on the West Coast, I log in.

First, I check updates on my posts. Most of them don't warrant interaction from me.

Diya Agarwal (real name, verified, like almost everyone on the site) critiqued a critique of a critique of a paragraph in my post. It's a solid critique. I complete-agree-vote the whole critique, and I weak-signal-boost-vote the sentence that contains the main minor insight to anyone who viewed the relevant paragraph in the OP. This only takes a few keystrokes and a few quick glances, because this was one of the top actions suggested by the Lilim that guessed what reactions I might want to give (and the UI team put a lot of labor and love into good scaffolding, clear presentation of actions, and easily user-configurable keyboard shortcuts). I then also accept most of the Lilim-suggested package of also weakly upvoting Diya's karma, credibility, and how much I defer to her, both in general and domain-specifically (except I remove the general-defer-vote, as I don't know how much to generalize before seeing more from her); and in the same keystroke I accept the top Lilim-suggested explanation for those upvotes. I make an additional warrants-further-investigation react, which the Lilim did not anticipate, and leave a quick comment explaining how Diya's point generalizes and suggesting some places it might have further implications. Users who were eigendeferring to me in this domain now slightly eigendefer to Diya.

Over the next few minutes, Diya and a few others metareact to my reactions. Most importantly, George Lark complete-disagree-votes and karma-downvotes my upvoting Diya's general and domain-specific karma. His stated reason (most interactions are required to have a stated reason, except e.g. complete-agree-votes which have a default meaning or some karma-upvotes where it's unambiguous why, e.g. your only manual input was interesting-voting and you accepted the Lilim-suggested consequents without review) is that his previous comment already made a point that has Diya's point as a natural implication. I complete-agree-vote his clause that claims his previous comment is prior art, and I partial-agree-vote and partial-disagree-vote (this is a single action, for roughly equal degrees of agree and disagree) his complete-disagree-vote on my karma-upvote of Diya; my explanation is that Diya actually definitely should get some credit, because the Lilim had not noticed the implication of George's comment for the context where Diya commented (and I had not seen George's comment), so Diya provided real additional work—but also George is correct that some of my upvote was for the insight itself, and so some of that karma should go to him for first publishing. (In the background, a Lilinn checks Diya's and George's read-logs, finding that Diya did not see George's comment (unless she is Circumventing) and did not have that much overlap with George's sources, so Diya gets substantial independent-discovery credit. Both of their read-logs are now indexed to this insight, for the Lilim to offer up in case a user asks for background text for that insight.) In two keystrokes I accept the Lilim suggestion to objectlevel-normbreak-vote his karma-downvote of my karma-upvote of Diya, and its proposed explanation that karma-downvotes are supposed to be about the user's judgement (me), not the action—George would have known that I didn't know that there was his prior work, at the time that I made the karma-upvote. I also metalevel-normbreak-vote his comment, with the explanation that I checked the logs of what the Lilim showed him when he was writing his comment—and I can see that he must have either completely ignored the Lilim's explanation of the relevant objectlevel-norms, or else saw it but disregarded it without acknowledgement, and in this case the Lilim correctly identified his norm mistake. It is a norm about norms that you should check the top Lilim warning about norms that your interaction might be breaking. George rejects the top Lilim suggestion of complete-disagree-voting my reactions, and instead accepts the Lilim-suggested package of complete-agree-voting my reactions, which a Lilinn takes as a cue to automatically (slightly) decrease his meta-karma-react credibility and increase his dispute-resolution-credibility.

I add quick speculations on how ego syntonicity might interact with voting patterns, which the Lilim correctly label as butterfly ideas. Since they are labeled that way, users know not to crush them. A Lilinn directs butterfly idea posts to users to read only if they are in the mood for those posts. If the Lilinn isn't sure whether the user is in the mood, it can check (by making a top-suggestion reaction-package say "I didn't evaluate this but I don't want to see it right now"), and usually home in on most of the main functional axes of the user's mood fairly quickly.

With that sordid upkeep business finished, I move to my second customary step. I open the auto-moderation logs, and take a quick look.

A few users have been given several-week bans for sustained poor behavior, earning them too much negative metalevel-norm-karma. (Users don't get banned for poor epistemics, they just get their posts deboosted—shown mainly to users who are interested in helping others improve their epistemics, with prominent warnings about overall and domain-specific epistemic karma, and links to specific mistakes.) This is sad but necessary, and while I do want to see the notification, I don't have anything to add. This is uncommon, but not a surprise—a small faction of users had been developing a coalition of alternative norm enforcement, and instead of discussing the merits and demerits of that alternative equilibrium, they were trying to push it into the normspace by a sum-threshold attack where the targeted pattern was too subtle for the Lilim to recognize de novo and was too inexplicit for individual targeted users to be easily group-conscious of, but real enough to impact the targeted users. It partly worked, in that the Lilim had mostly not noticed, except that they flagged some of it as a possible brigading pattern (statistically ambiguous with natural correlations of voting judgements and content interests) among many other possible brigading patterns. But the users noticed eventually, tutored the Lilim to recognize the attack pattern and promote it to user attention as a hypothesis with other context from the attacking user, and accordingly metalevel-norm-karma downvoted (which becomes very easy when you can just accept the explanation given by another trusted user) the attackers (who did not respond to discussion offers).

A couple new visitors have flamed out. This, unfortunately, is not uncommon. Most users, numerically, who land on a page on the forum trying to read are either confused, or fascinated and happily read more. Most users, numerically, who try to engage in the forum run away immediately (screaming, presumably) with no explanation. We have been A/B testing gentle ways to introduce the forum, including having new-user regions, Lilim tutors and assistants, a curriculum, and so on, but... the forum is just not for most people, and that is ok. One of the flame-outs was flagged by the post-mortem summarization Lilinn as being potentially interesting and good-faith. I glance at the logs. They were not accustomed to having everything, including the Lilim help chats, be public; and the sheer complexity of the UI, even in its very simplified intro form; and the apparently extreme levels of hyperjudgemental hypermonitoring and credit scores; and the seemingly meaningless and definitely incomprehensible navel-gazing meta-meta-norm fights. But yes, they did seem to have a spark of important thinking about things that matter. Feeling a bit unusually extraverted, I shoot them an email offering to chat live and in private (and I log that I've done so, to prevent users spamming them).

My third customary step is to do my due diligence, and visit a couple other of the major sectors of norm-system-space, mostly at random. A Lilinn suggests a few points that might be interesting to me. They are interesting, and I write some follow-on thoughts that are visible in my home sector, but I quickly have to block notifications, and deboost my post in the other sector, and medium-strength suggest a blanket temporary quarantine shadowblock of any influx from the other sector into my home sector coming through my reaction post.

Those people's behavior is just too costly to deal with and not worth-it enough. They reliably crush butterfly ideas; they punish people for making statements that are correlated with political positions they don't like; they won't feel an obligation to engage in meta-discourse or even state justifications for reactions(!); they allow and enact meta-react blocks; and they don't think they shouldn't do those things, so they don't enforce norms against those things; and they punish people who try to enforce those norms, or to discuss whether those norms should be enforced. And, may God eventually forgive them, they will go around karma-downvoting unrelated posts if they dislike something you said! Trying to bridge the metalevel-norm gap is a noble and worthwhile task, but it's... a task, a multi-day—nay, multi-year—task that can't be automated, is twisty and fragile and recursive, and just isn't likely to be high-priority for this specific instance; the Lilim know to un-quarantine the strangers for any home users who are explicitly trying to study this phenomenon or bridge the specific gap. Almost all home users defer to my quarantine suggestion (almost all users defer to almost all users's quarantine suggestions). I retreat to my sector where it's possible to think.

Fourth: Let's browse.

I click a couple posts to read. One of them is long, and is a mix of stuff I do and don't already know, so I talk to a Lilinn about it, asking basic questions about what the author is saying. Like most serious users, I have my reader Lilinn scaffolded to mainly give very short conservative glosses, and to point at the relevant few paragraphs; I only rely on glosses when the Lilim are confident that the gloss is basically lossless, e.g. because the paragraph being glossed is not a new idea or a new expression, but is just a report of the author's particular position within an already well-understood space of well-understood positions. (For basic reading, a single Lilinn is most of what's required, because this is such a common task and pretty easy; but substantial glosses that the reader Lilinn decides to show to me are passed through a fuller host of Lilim to verify and confidence-tag and maybe context-fetch.)

The author of this post (who, in a rare exception, is anonymous, having been verified privately with the moderation team, usually due to living under a repressive regime) has pretty high general-credibility-karma and domain-credibility-karma according to the consensus-centrality-weighted-home-sector-population judgement; and furthermore has high general- and domain-credibilities according to the eigenjudgements of several of my various moods and modes. (My current mood could be called open-speculative-sociological-foraging-selfdirected, for which the author has high karma; the author also has high credibility for my focused-pruning-textual mood; the author has fairly low karma for my debate-clarification-avantgarde-canonical-leader mood, but I don't feel like that right now.) For this reason, I'm indirectly eigendeferring to this author on this topic. That's part of why this post is getting so much attention; this author is eigendeferred-to on this topic by a substantive chunk of our home sector (and even many users from other sectors, though that carries much less weight). Added to that, is the fact that, as far as the LILIL ensemble can tell, if one uses quasi-logical deduction to construct quasi-coherent theories out of the myriad eigenopinions floating around, one gets a few dozen overarching (overlapping, still partial) Perspectives; and the topic of this post is a meso-sized subpillar of a somewhat widely subscribed (literally, as in having been directly or indirectly partially signed off on, and also in terms of readership) Perspective; and this author is one of the main eigendeferrees for this Perspective on this topic; and therefore this is an important discussion worthy of eyeballs.

As I engage with the post, I do nod along with a lot of it, mainly deferring; but I notice a couple threads of subtle but important disagreement, or at least confusion—I detect something is off / unresolved / incomplete / anomalous. I note this in the form of disagree-votes and some reacts such as "I feel vaguely bad or uneasy about this statement", but I do so within a deferential modal-setting. That means my reactions are largely not counted for object-level aggregation, e.g. a given paragraph doesn't accumulate much direct disagreement from my reaction; and the Lilim won't try very hard to interpret my reactions as coming from an alternative coherent viewpoint (unless it just indexes to an existing known viewpoint); but my reactions do affect eigendeference strengths. In particular, my reactions, combined with similar reactions from several other users, and the collection of users who defer to us in this context—as well as, it turns out, some symmetric reactions to the opposing Perspective's eigenopinion on this topic—causes a substantial increase in the fluidity of these Perspectives around this topic and a few topics that are dependent on this topic. This increases the salience of investigation (fact-finding, theorizing) around this topic, and weakens several eigenopinion-networks. Truly exciting times.

Fifth and finally, it's time to create for the day. I settle down to write.

In the course of my investigations and record-making, I have a long phrase that I have to deploy three times. I get piqued, and ask the Lilim to find words / grammar for this phrase. It fails after twenty seconds. I ask it to try harder to find examples of the same function in other texts. It fails after fifteen minutes, just turning up two kinda-examples that don't really help and several non-examples. I give more of a prompt, and ask the LILIL ensemble. It asks more clarifying questions to pin down what I'm looking for in contrast to the Lilim's attempted examples, and it searches very widely, promoting many texts to its attention for any one of many reasons (keywords, searches on cached hidden activations, reevaluating anew, rewriting in several variations and debating whether the rewrites are accurate, random sampling, following citation graphs locally or in the direction of users whose thinking has various substantial dot-products with mine, asking itself what other strategies to use) and then debating amongst itself whether the text might be a subtle but true example. It returns a few dozen attempts, most of which are incorrect, but ten of which are close. Four among those are true examples. The examples stimulate me to crystallize, generalize, and factor into two axes and a few different concepts. I put out a request-for-word with the full context, for the LILIL ensemble and the community to work on in the background.

As I finish up the bulk of my investigation for the day, I contact one of the Living Libraries who specializes in the general area, to test my current theorizing. These users have extensive knowledge of the Corpus of the forum. They mainly don't do their own iterative-theorizing-investigation, though they do write a fair amount—helping with lexicogenesis, scaffolding the LILIL ensemble for difficult tasks in finding texts (which will then be used to scaffold and fine-tune Lilim), organizing Perspectives, editing Wiki entries. Instead, they read immense quantities from the Corpus, including old and obscure text, categorizing and writing short indexing-summaries, and reading each other's index-summaries. Sometimes they compile sets of texts to make a request-for-refactoring or a request-for-distillation to the community. I share some of my thoughts and quickly note the related work that Lilim brought up, and how my thinking seems to be going off in a different direction. The Library points me at a few key sources, one of which seems quite parallel—or fruitfully parallax—to my lines of thinking. I reach out to the author for a synergistic collaboration.